The Invoice Always Comes After the Magic Show

Gartner says at least half of all AI projects are going to blow their budgets. I think that's optimistic. I think the real number is higher, because the people who are going to catch the problem are the ones who have to admit they didn't see it coming, and nobody wants to be the person who sits across from finance and says "our AI strategy is failing because we forgot to calculate the cost of inference." So they'll kick it down the road, add another zero to the budget, and hope the technology gets cheaper before the board gets curious.

The thing about AI, about any new technology really, is that it's sold as a vision, a promise of transformation, of revolution, of magic (and who doesn't love magic, who doesn't want to believe there's a way to make all the hard things easy, all the complicated things simple), but the reality is always more expensive than the demo, because the people selling the technology are not the people running it. The people running it are the ones who get to have the conversation with finance about why the pilot that cost ten thousand dollars now costs a million, and that conversation is never fun, that conversation is always about cutting costs, about reducing the quality of the output, or reducing the people who were supposed to oversee it. The engineers who warned about this get laid off in the reorg. The people who greenlit the project get promoted. The pattern is so old it's boring, but it keeps happening because we keep pretending that technology is magic instead of infrastructure.

We did this with the cloud. We did this with microservices. We did this with blockchain. We're doing it with AI, and the specifics change but the shape stays the same: someone sells a vision that is technically true in a very narrow set of circumstances, and then the people who have to make it work in the real world get to explain why the vision was always going to hit a wall made of operational reality. I'm not anti-AI. I write code, I use tools, I think there are useful things here. But the dishonesty of the selling is corrosive. The pitch is always "transform your business" and the reality is always "here is a new category of recurring expense you did not budget for," and somewhere between those two sentences is where the project dies, or gets restructured, or gets handed to a new team with a new mandate and the same broken assumptions baked into the architecture.

There's a human cost in here that the Gartner report doesn't touch, because reports don't touch human costs, reports touch percentages. When these projects overrun, the response is rarely "let's be more thoughtful about what we're building." It's "cut costs," and cutting costs in AI usually means either degrading the output (which makes the tool worse) or cutting the people who were supposed to oversee it (which makes the problems worse and just delays the moment of reckoning to when there's nobody left who remembers why the decision was made). The magic is real until the invoice comes. And then the magic is just a line item that someone has to explain, every call, every token, every time the agent loops back to check its work adding to a number that nobody modeled in the original proposal because the original proposal was built around a demo that ran for twenty minutes in a controlled environment while everyone in the room was still excited.

The real lesson isn't about AI. It's about the difference between building a prototype and running a service. A prototype is a promise. A service is a commitment. Commitments have costs that promises don't, ongoing costs, operational costs, the kind that compound quietly in the background until someone pulls a report and the number is a lot bigger than anyone thought it would be by now. I don't know what happens when half these projects blow their budgets. Maybe the industry course-corrects and we get smarter about architecture. Maybe the hype deflates and the people who actually know what they're doing get to build something sustainable without a VP breathing down their necks about the demo timeline. Or maybe we just do what we always do: throw more money at it, pretend the numbers will work out eventually, and hand the mess to the next team with a fresh mandate and no institutional memory of why the last team left.